Transcription Models
Local and cloud transcription models — Whisper, Parakeet, and cloud providers.
VivaDicta supports a wide range of transcription engines — both local models that run entirely on your iPhone or iPad and cloud providers that offer speed and accuracy.
Local Models
Local models process audio directly on your device. No data leaves your iPhone, and they work without an internet connection.
- Whisper — OpenAI's Whisper model optimized for Apple hardware. Multiple model sizes available (we recommend Large Turbo). Best balance of accuracy and speed on modern iPhones.
- Parakeet — NVIDIA's speech recognition model running via FluidAudio. Fast and accurate, optimized for Apple Silicon.
Cloud Providers
Cloud providers process audio on remote servers. They tend to be faster (especially on older devices) and often more accurate, but require an internet connection and an API key.
Some cloud models also stream: instead of uploading the finished recording and waiting, they transcribe over a live connection while you speak, so the text is ready almost the moment you stop. See Real-Time Transcription below.
- Groq — free forever, ultra-fast Whisper on custom LPU hardware. Our #1 recommendation.
- Deepgram — Nova-3 model with excellent accuracy. $200 free credits for new accounts.
- ElevenLabs — Scribe v2 with support for 99 languages. Free tier available.
- Gemini — Google's multimodal model with speech transcription capabilities.
- Mistral — European AI provider with transcription support.
- Soniox — high-accuracy cloud transcription.
- OpenAI-compatible — connect any provider that supports the OpenAI Whisper API format.
Local vs Cloud
| Local | Cloud | |
|---|---|---|
| Cost | Free | Free tiers or pay-per-use |
| Privacy | Full — nothing leaves your device | Audio sent to provider servers |
| Internet | Not required | Required |
| Speed | Depends on device hardware | Consistently fast |
| Accuracy | Good to excellent | Excellent |
| Setup | Model download required | API key required |
Speaker Diarization
Some models can identify and label different speakers in a multi-voice recording, producing a speaker-separated transcript. Turn this on in Settings > Transcription > Speaker Labels - the setting only takes effect when your active model supports it.
| Model | Speaker Labels |
|---|---|
| Whisper (local) | Yes |
| Deepgram | Yes |
| Mistral | Yes |
| Other models | Not supported |
See Speaker Labels for how to enable the feature and tips for cleaner results.
Language Support
Language support varies by model. Most models support 50-100+ languages. Whisper (local) and Groq (cloud, also using Whisper) support 100+ languages. ElevenLabs Scribe v2 supports 99 languages.
See Recommended Models for our top picks based on your use case.
Real-Time Transcription
Most models work by upload: you stop recording, the audio goes to the provider, and you wait for the transcript to come back. Streaming models work while you talk, so by the time you stop, only the last few words are outstanding.
These models stream:
- Soniox Realtime — 60+ languages, the broadest language coverage of the streaming options.
- Deepgram Nova 3 — the same model as the upload path, served over a live connection.
- Deepgram Flux — streaming only, in English or multilingual builds.
- ElevenLabs Scribe — streams the same Scribe model you would otherwise upload to.
- Voxtral Realtime — Mistral's streaming model.
Streaming works both inside the app and from the keyboard. If the connection drops at any point, VivaDicta quietly falls back to uploading the recording it was writing all along, so a lost connection costs you speed, never the transcript.
Two settings turn it off for a given mode. Speaker Labels and inline translation both need the whole recording before they can work, so a mode using either takes the upload path instead.